88% of YC startups are AI-based. The real revolution isn't the models
88% of YC startups are AI-based. But the real revolution isn't in the models at all.
I was sure I was on top of the AI industry. Read Twitter, tracked the big launches, tested new tools. Felt like I had the context. Then I spent four days taking apart the companies from the latest Y Combinator batches. 100+ AI startups that raised rounds. Looked at MRR, team size, who the founders are, what exactly they do. What problem they solve. Dumped it all into a spreadsheet. Analyzed it...
And realized I hadn't actually understood what's going on.
First, the numbers: across the 2025 batches, up to 88% of startups are AI-native. Half are building agents. 68% are B2B. A quarter of the W25 batch wrote 95% of their code with AI tools. Companies reach $10M revenue with teams under 10 people.
But that's not what hooked me. It's the trends, and they're impossible to miss:
1/ Vertical agents ate the horizontal ones
YC no longer takes companies building AI assistants for everything. Companies differentiate inside specific industries. Often very narrow ones:
- LunaBill. A voice bot for clinic billing. One job: calling insurance companies back about unpaid claims. Launched in July; by year's end, $428K in real revenue plus another $336K in signed contracts. Every client who tried the pilot became a paying one. Every single one! The old me would have said: "It's just a wrapper, nothing of their own." But that's exactly why it's working: they don't sell technology. They solve a specific pain, in a very narrow industry.
- Vesence. The founders spent months living inside a Swedish law firm. Built a copilot that checks documents for mistakes before they go out to the client. First firm-wide rollout: 90% of employees use it every week. Raised a round to hire engineers and pay for LLM calls. Their biggest expense line is literally prompts 🙂
2/ Agent infrastructure is the new picks and shovels
- When everyone is building agents, someone has to build infrastructure for the people building agents. Memory layers, observability, MCP gateways. Systems for grading answers (evals). Y Combinator is actively stocking up on this layer, all through the Fall 2025 intake.
Stripe for payments. AWS for servers. The same kind of invisible layer is forming for agents now: nothing works without it, and everything stands on it. And this is where the strongest lock-in lives: once a company starts building agents on your infra, leaving you is very expensive.
3/ 10 people. $10M revenue.
This number won't let me go. Garry Tan (CEO of Y Combinator) says the next $100B companies will be 10 people each. The defining metric of the era: revenue per employee.
- VideoGen: two founders straight out of a dorm room. Starting capital: $30K saved from internship paychecks at Palantir and AWS. Bootstrapped to $1M+ ARR before YC. Now they serve teams at Google and ByteDance. This used to be the exception. Now it's the baseline, and a quarter of the batch.
TL;DR:
1/ Stop watching the top layer. Model launches and benchmarks are hype. The real action is one level down — in boring niches, where tiny teams build products for very clear jobs.
2/ Models are slowly turning into a commodity. The differentiation comes from deeply understanding a narrow domain, the technology, and how to use both to solve problems for people and businesses.
P.S. If you want the list of companies I analyzed, ping me in the dm